Public opinion polls in the 2020 presidential election suffered from errors of “unusual magnitude,” the highest in 40 years for surveys estimating the national popular vote and in at least 20 years for state-level polls, according to a study conducted by the American Association for Public Opinion Research (AAPOR).
Polls understated the support for Trump in nearly every state and by an average of 3.3 percentage points overall. Polls in Senate and gubernatorial races suffered from the same problem.
“There was a systematic error that was found in terms of the overstatement for Democratic support across the board,” said Josh Clinton, a Vanderbilt University political science professor who chaired the 19-member task force. “It didn’t matter what type of poll you were doing, whether you’re interviewing by phone or Internet or whatever. And it didn’t matter what type of race, whether President Trump was on the ballot or was not on the ballot.”
The polls did a better job of estimating the average support for Biden, with a few exceptions. In general, support for Biden in the polls was 1 percentage point higher than his actual vote.
An AAPOR task force conducted a similar examination after the 2016 election. Then, national polls generally accurately predicted the size of Hillary Clinton’s popular vote victory over Trump, but state polls proved more problematic, causing many analysts at the time to predict wrongly that Clinton also would win an electoral college majority.
In the new study, task force members were able to rule out a series of reasons that might have caused the 2020 polls to show a bigger margin for Biden over Trump than the actual results. That included some of the problems that affected polling in 2016, such as the failure in that year to account for levels of education in the samples of voters.
But the task force members were not able to reach definitive conclusions on exactly what caused the problems in the most recent election polls and therefore how to correct their methodology ahead of the next elections. “Identifying conclusively why polls overstated the Democratic-Republican margin relative to the certified vote appears to be impossible with the available data,” the report states.
Polling in senatorial and gubernatorial races showed a similar pattern, overstating the margin for Democratic candidates versus their Republican opponents. When state-level presidential polls were removed from the sample, the error level was even higher. For example, polling pointed to possible Democratic gains in House races. Instead, Republicans gained seats.
National presidential polls were accurate in one respect, which was in pointing to the popular vote winner. Of 66 such polls taken in the last two weeks of the campaign, all showed Biden ahead of Trump. “The performance of senatorial polls was notably worse,” the report notes, with just 66 percent correctly identifying the winning candidate.
Josh Clinton, the task force chairman, said the group ruled out some possible reasons the polls were not as accurate as hoped. For example, the proposition that some Trump voters lied to pollsters about how they would vote doesn’t hold up because the error margin was larger for Republicans in Senate and gubernatorial races.
The task force examined and ruled out sources of error in 2016 polls as the cause of problems in 2020. In 2016, late-deciding voters broke heavily for Trump. In 2020, there were far fewer such voters, with most voters generally locked in on their choice well before Election Day and with many casting ballots in advance, either by mail or in person.
Another factor that contributed to problems in 2016 was that many polls did not weight their samples based on education levels, a relatively new fault line in voting behavior. The 2016 election saw clear differences in vote preferences among White voters with college degrees, who generally favored Clinton, versus those without, who generally favored Trump. By 2020, most polls did weight their samples by levels of education.
Nor was there evidence that the polls were mistaken in their assumption of the composition of the electorate. No group or groups were systematically underrepresented or overrepresented in the pre-election polls, the report said.
The task force examined other possible causes for error, such as whether supporters of Trump and Biden told pollsters how they would vote but ultimately did not vote or whether polls included too many people who voted early (a group that favored Biden) and too few who voted on Election Day (a group that favored Trump). In neither case was that shown to be a problem.
One possible explanation is that Republicans who responded to surveys voted differently than Republicans who choose not to respond to pollsters. The task force said this was a reasonable assumption, given declining trust in institutions generally and Trump’s repeated characterizations of most polls by mainstream news organizations as fake or faulty.
“That the polls overstated Biden’s support more in Whiter, more rural, and less densely populated states is suggestive (but not conclusive) that the polling error resulted from too few Trump supporters responding to polls,” the report states. “A larger polling error was found in states with more Trump supporters.”
The report points to the role of new voters in 2020 as one possible cause of the errors. There were 22 million more votes certified in 2020 than in 2016, enough in some states to account for the error rate in the polls. But the report said that the available data is not sufficient to draw conclusions about whether these new voters were the main source of polling errors.
The report emphasizes that though often quite accurate, polls are not as precise as sometimes assumed and therefore given to misinterpretation, especially in the most competitive races. “Most pre-election polls lack the precision necessary to predict the outcome of semi-close contests,” the report states. “Despite the desire to use polls to determine results in a close race, the precision of polls is often far less than the precision that is assumed by poll consumers.”
The report offers a cautionary note about the reliability of pre-election polls in the upcoming elections, stating, “It is unclear whether the problems polls faced in 2020 will persist in 2022 or 2024, and what happens in 2022 may be uninformative for knowing if there are longer-term issues.”
Clinton pointed to the Trump factor in the context of future elections and polls. “It’s possible that if President Trump is no longer on the ticket or if it’s a midterm election where we know that the electorate differs in the presidential election, that the issue will kind of self-resolve itself,” he said.
He said that polling in 2018 was generally better than in 2016, which led some pollsters to believe the problems had been resolved. Then came 2020, and the problems reemerged. That has left a high degree of uncertainty as to whether the problem is mostly related to whether Trump is on the ballot or whether there are deeper problems affecting polls.
Clinton said that, if polls in 2022 are not particularly accurate, that would be a sign of a persistent shift in pollsters’ ability to reach particular groups of voters. “But if the polls do well in 2022, then we don’t know if the issue is solved,” he added. “Or whether it’s just a phenomenon that’s unique to presidential elections with particular candidates who are making appeals about ‘Don’t trust the news, don’t trust the polls’ that kind of results in taking polls becoming a political act.”